Dictating a minute of notes after leaving a pharmacy can save you from reconstructing the visit that evening. The delicate step comes next: turning “considering twelve units” into useful follow-up information without creating an order the customer never confirmed.
This method starts with dictating your own notes after the appointment in a company-approved tool. Recording the customer conversation is a different use that needs its own legal and organisational assessment. You do not need a conversation recording for this trial.
Dictate the facts while they are still clear
Stop somewhere away from the shop’s conversations. State the outlet ID, the date and the details that change the next action. Distinguish what you observed, what the contact said and what remains unconfirmed.
Say the units and qualifications aloud: six individual items are not six cases, and visible stock is not the shop’s total stock. Make spoken corrections explicit. “Twelve—no, six units observed” should result in six units, with the transcript available for review.
Turn the note into fields you can check
| Source passage | Proposed field | Status |
|---|---|---|
| Six visible units; stockroom not checked | Shelf observation: 6 units of X | Total stock unknown |
| Considering twelve; nothing confirmed | Interest in 12 units of X | Proposal to discuss; no order |
| Send price list on 11 September | Task: send price list, due 11 September 2026 | For the rep to complete |
| Call back on 15 September | Task: call, due 15 September 2026 | To schedule |
| Discount needs internal approval | Commercial approval request | No discount granted |
Section 2.2 of the NIST Generative AI Profile describes confidently stated false answers. For a visit record, the risk can lie in a single word: dropping “not” turns a refusal into agreement. A fluent summary does not replace checking fields that trigger an action.
Review before writing to the CRM
The workflow can stay simple: audio note, transcript, proposed fields, review, then save. Put the transcript beside the suggested fields. The rep should be able to correct a quantity or remove a task without entering everything again.
- →Identity: attach the note to the selected outlet ID, not just a matching name.
- →Products: check the SKU and pack size in the catalogue; do not automatically substitute a similar reference.
- →Decisions: preserve the difference between a request, an intention and a confirmed agreement.
- →Dates: interpret “tomorrow” from the local dictation date, not the date of a delayed sync.
- →Owner: assign the action to an identified person or ask for confirmation if the note does not specify one.
An integration must respect the rep’s permissions and approval process. Permission to create draft notes does not imply permission to approve orders or change pricing. Agree those boundaries with the team administering the CRM, whichever assistant you choose.
- Dictate
Your own notes after the visit, including outlet and date.
- Transcribe
Preserve units, negatives and spoken corrections.
- Structure
Suggest fields alongside the supporting passages.
- Review
Read, correct and save the note and tasks.
“Considering twelve units” remains an intention. This workflow does not approve orders.
Handle lost connections without duplicating tasks
If dictation is stored offline, show clear states: “saved on this device”, “uploaded”, “ready for review” and “saved in CRM”. The rep needs to know what has completed before deleting the note or ending the day.
For the integration, assign a stable identifier to each note and its approved version. Two attempts to send the same approved version should result in one record and the same tasks. This idempotency rule prevents duplicate reminders after a network interruption. You still need acknowledgement: sending a request does not prove that the CRM accepted the note.
Test errors with a commercial consequence
Build a trial set of fictional or authorised notes: street noise, similar product names, a spoken correction, no agreement, a relative date and duplicate synchronisation. Check the transcript and then the fields so you know where the workflow failed.
Example: 20 notes with five checked fields each give 100 checks. If 92 are correct, the accuracy is 92% of fields, not 92% of notes. This percentage also hides severity. One order created without agreement is enough to suspend that automatic write, even if names and dates are almost always recognised.
Time the complete process. If manual entry takes six minutes, compared with one minute of dictation and two minutes of review, the fictional saving is three minutes per note, or 36 minutes for 12 visits. Include failures and training time before expanding the trial.
Keep only what professional follow-up requires
The GDPR principles explained by the CNIL (French) call for a defined purpose, limited data, appropriate access and retention rules. Exclude patient names, purchases and health information from this workflow. Define who can listen to the audio and when it is deleted; its retention period need not match that of the approved sales report.
Start with your visit report template
Bring a fictional visit and the fields your team uses. We can examine the data-entry and follow-up steps during the demonstration.
Next, try a follow-up drafted from the approved note as part of your field-sales AI pilot. Our visit report template (French) and EDI and API guide (French) cover the underlying workflow.
By Anatole Briant
By Pierre Lerebours
By Anatole Briant
COO of Salesia
Pierre Lerebours is COO of Salesia. He is your contact for implementation, training and ongoing support.